Guides · Updated
Choosing a mini PC for local AI
How to choose a mini PC for local models: a basic small PC, a Mac mini or Mac Studio, or a unified-memory AMD machine, and what each one can run.
“Mini PC” covers three quite different computers. The right one depends on whether you want a quiet box for small models, a simple Mac, or enough memory to load a very large model. Which models fit still comes down to memory, which VRAM and quantisation explained covers.
The three kinds
| Kind | Choose it when | Watch out for |
|---|---|---|
| Basic mini PC (a typical small Intel or AMD office box) | You want a small, low-power machine for small models and ordinary desk work | No strong GPU. Expect the same limits as a PC with no graphics card. |
| Mac mini or Mac Studio | You want a quiet machine that runs Ollama and LM Studio on the GPU with no extra drivers | Memory is fixed when you buy it. Base chips have less memory bandwidth than Pro, Max and Ultra chips. |
| Mini PC with AMD Ryzen AI Max+ 395 (shared memory, up to 128 GB) | You want to load models that do not fit on a normal graphics card, in a small box | Shared memory is slower than a desktop card’s VRAM, so a large dense model can feel slow even when it fits. Ollama lists this chip for ROCm on Linux. |
Driver detail for the AMD machines, and for Intel Arc, is in Running local AI on AMD, Intel Arc and mini PCs. Mac setup is in Local AI on a Mac with Apple Silicon.
What to check before you pay
- How much memory, and can you add more later? On a Mac, and on a Ryzen AI Max+ 395 machine, the memory is chosen at purchase. Buy the amount you want for the models you care about. A basic mini PC often takes laptop-style memory modules. Confirm that on the spec sheet. If it has two slots, fit a matching pair so both memory channels are active. Why that matters: Local AI on a budget PC.
- A large model, or a fast reply. Unified memory is how a small box holds a large model. A graphics card in a normal desktop answers faster. If you have room for a tower and speed is the priority, a discrete GPU is the better tool. Compare tiers on the hardware pages.
- The NPU rating. “AI PC” numbers mostly describe a neural processing unit. Ollama and the other chat apps in these guides run on the GPU and the CPU. Memory is the spec that decides which models fit.
- Disk space. Each model is several gigabytes, and a few of them add up. Look for an SSD with room to spare, and check whether a second drive will fit. Moving models later: Keeping local models up to date and tidy.
- Cooling. A reply can keep the chip busy for a long stretch. A fanless box may slow itself down as it warms up. A Mac mini, a Mac Studio, or a mini PC with a proper cooler holds its speed better. The same point for laptops is in Running local AI on a laptop.
- The operating system. A Mac uses the GPU through Metal. For a Ryzen AI Max+ 395, Ollama’s ROCm support is listed for Linux. On Windows, check Ollama’s hardware page before you assume the fast path is available. Vulkan covers a wider set of AMD and Intel GPUs, with less speed than ROCm on a supported card.
What to run on each
- Basic mini PC: small models, the same starting points as Local AI on a budget PC.
- Mac: follow the memory table in the Mac guide. 16 GB is a sensible minimum if you are buying one for this.
- Ryzen AI Max+ 395 with a large memory pool: mixture-of-experts models, which read only part of their weights for each word. Suitable tags include
qwen3.6:35b,gemma4:26b,gpt-oss:120bandqwen3-coder:30b. More on that class of machine: AMD, Intel Arc and mini PCs.
For a machine you already own, use the GPU picker. Choose Apple Silicon Mac or No GPU / CPU only where that applies.
When a mini PC is the wrong purchase
You already have a desktop. More RAM, or a graphics card with more VRAM, usually changes what you can run more than a second computer does. The upgrade order is in the budget PC guide.
Image generation leans on a discrete GPU much more than chat does. See Generating images locally.
After it arrives
Install Ollama, pull one small model, and run ollama ps. On a Mac, or on a supported AMD box, the PROCESSOR column should show the GPU in use. Then try the next model up.
Suggested machines are on our mini PC hardware page. As an Amazon Associate, Lila Forge earns from qualifying purchases. Links on that page go to Amazon and are marked rel="sponsored nofollow". Buying through them costs you nothing extra, and it doesn’t change which machines we list.
A small office sharing one box: Local AI for a small business: a practical starter.
Sources: Ollama hardware support, Ollama Windows docs.
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →